File: /usr/src/linux/arch/sparc/math-emu/math.c
1 /*
2 * arch/sparc/math-emu/math.c
3 *
4 * Copyright (C) 1998 Peter Maydell (pmaydell@chiark.greenend.org.uk)
5 * Copyright (C) 1997, 1999 Jakub Jelinek (jj@ultra.linux.cz)
6 * Copyright (C) 1999 David S. Miller (davem@redhat.com)
7 *
8 * This is a good place to start if you're trying to understand the
9 * emulation code, because it's pretty simple. What we do is
10 * essentially analyse the instruction to work out what the operation
11 * is and which registers are involved. We then execute the appropriate
12 * FXXXX function. [The floating point queue introduces a minor wrinkle;
13 * see below...]
14 * The fxxxxx.c files each emulate a single insn. They look relatively
15 * simple because the complexity is hidden away in an unholy tangle
16 * of preprocessor macros.
17 *
18 * The first layer of macros is single.h, double.h, quad.h. Generally
19 * these files define macros for working with floating point numbers
20 * of the three IEEE formats. FP_ADD_D(R,A,B) is for adding doubles,
21 * for instance. These macros are usually defined as calls to more
22 * generic macros (in this case _FP_ADD(D,2,R,X,Y) where the number
23 * of machine words required to store the given IEEE format is passed
24 * as a parameter. [double.h and co check the number of bits in a word
25 * and define FP_ADD_D & co appropriately].
26 * The generic macros are defined in op-common.h. This is where all
27 * the grotty stuff like handling NaNs is coded. To handle the possible
28 * word sizes macros in op-common.h use macros like _FP_FRAC_SLL_##wc()
29 * where wc is the 'number of machine words' parameter (here 2).
30 * These are defined in the third layer of macros: op-1.h, op-2.h
31 * and op-4.h. These handle operations on floating point numbers composed
32 * of 1,2 and 4 machine words respectively. [For example, on sparc64
33 * doubles are one machine word so macros in double.h eventually use
34 * constructs in op-1.h, but on sparc32 they use op-2.h definitions.]
35 * soft-fp.h is on the same level as op-common.h, and defines some
36 * macros which are independent of both word size and FP format.
37 * Finally, sfp-machine.h is the machine dependent part of the
38 * code: it defines the word size and what type a word is. It also
39 * defines how _FP_MUL_MEAT_t() maps to _FP_MUL_MEAT_n_* : op-n.h
40 * provide several possible flavours of multiply algorithm, most
41 * of which require that you supply some form of asm or C primitive to
42 * do the actual multiply. (such asm primitives should be defined
43 * in sfp-machine.h too). udivmodti4.c is the same sort of thing.
44 *
45 * There may be some errors here because I'm working from a
46 * SPARC architecture manual V9, and what I really want is V8...
47 * Also, the insns which can generate exceptions seem to be a
48 * greater subset of the FPops than for V9 (for example, FCMPED
49 * has to be emulated on V8). So I think I'm going to have
50 * to emulate them all just to be on the safe side...
51 *
52 * Emulation routines originate from soft-fp package, which is
53 * part of glibc and has appropriate copyrights in it (allegedly).
54 *
55 * NB: on sparc int == long == 4 bytes, long long == 8 bytes.
56 * Most bits of the kernel seem to go for long rather than int,
57 * so we follow that practice...
58 */
59
60 /* TODO:
61 * fpsave() saves the FP queue but fpload() doesn't reload it.
62 * Therefore when we context switch or change FPU ownership
63 * we have to check to see if the queue had anything in it and
64 * emulate it if it did. This is going to be a pain.
65 */
66
67 #include <linux/types.h>
68 #include <linux/sched.h>
69 #include <linux/mm.h>
70 #include <asm/uaccess.h>
71
72 #include "sfp-util.h"
73 #include <math-emu/soft-fp.h>
74 #include <math-emu/single.h>
75 #include <math-emu/double.h>
76 #include <math-emu/quad.h>
77
78 #define FLOATFUNC(x) extern int x(void *,void *,void *)
79
80 /* The Vn labels indicate what version of the SPARC architecture gas thinks
81 * each insn is. This is from the binutils source :->
82 */
83 /* quadword instructions */
84 #define FSQRTQ 0x02b /* v8 */
85 #define FADDQ 0x043 /* v8 */
86 #define FSUBQ 0x047 /* v8 */
87 #define FMULQ 0x04b /* v8 */
88 #define FDIVQ 0x04f /* v8 */
89 #define FDMULQ 0x06e /* v8 */
90 #define FQTOS 0x0c7 /* v8 */
91 #define FQTOD 0x0cb /* v8 */
92 #define FITOQ 0x0cc /* v8 */
93 #define FSTOQ 0x0cd /* v8 */
94 #define FDTOQ 0x0ce /* v8 */
95 #define FQTOI 0x0d3 /* v8 */
96 #define FCMPQ 0x053 /* v8 */
97 #define FCMPEQ 0x057 /* v8 */
98 /* single/double instructions (subnormal): should all work */
99 #define FSQRTS 0x029 /* v7 */
100 #define FSQRTD 0x02a /* v7 */
101 #define FADDS 0x041 /* v6 */
102 #define FADDD 0x042 /* v6 */
103 #define FSUBS 0x045 /* v6 */
104 #define FSUBD 0x046 /* v6 */
105 #define FMULS 0x049 /* v6 */
106 #define FMULD 0x04a /* v6 */
107 #define FDIVS 0x04d /* v6 */
108 #define FDIVD 0x04e /* v6 */
109 #define FSMULD 0x069 /* v6 */
110 #define FDTOS 0x0c6 /* v6 */
111 #define FSTOD 0x0c9 /* v6 */
112 #define FSTOI 0x0d1 /* v6 */
113 #define FDTOI 0x0d2 /* v6 */
114 #define FABSS 0x009 /* v6 */
115 #define FCMPS 0x051 /* v6 */
116 #define FCMPES 0x055 /* v6 */
117 #define FCMPD 0x052 /* v6 */
118 #define FCMPED 0x056 /* v6 */
119 #define FMOVS 0x001 /* v6 */
120 #define FNEGS 0x005 /* v6 */
121 #define FITOS 0x0c4 /* v6 */
122 #define FITOD 0x0c8 /* v6 */
123
124 #define FSR_TEM_SHIFT 23UL
125 #define FSR_TEM_MASK (0x1fUL << FSR_TEM_SHIFT)
126 #define FSR_AEXC_SHIFT 5UL
127 #define FSR_AEXC_MASK (0x1fUL << FSR_AEXC_SHIFT)
128 #define FSR_CEXC_SHIFT 0UL
129 #define FSR_CEXC_MASK (0x1fUL << FSR_CEXC_SHIFT)
130
131 static int do_one_mathemu(u32 insn, unsigned long *fsr, unsigned long *fregs);
132
133 /* Unlike the Sparc64 version (which has a struct fpustate), we
134 * pass the taskstruct corresponding to the task which currently owns the
135 * FPU. This is partly because we don't have the fpustate struct and
136 * partly because the task owning the FPU isn't always current (as is
137 * the case for the Sparc64 port). This is probably SMP-related...
138 * This function returns 1 if all queued insns were emulated successfully.
139 * The test for unimplemented FPop in kernel mode has been moved into
140 * kernel/traps.c for simplicity.
141 */
142 int do_mathemu(struct pt_regs *regs, struct task_struct *fpt)
143 {
144 /* regs->pc isn't necessarily the PC at which the offending insn is sitting.
145 * The FPU maintains a queue of FPops which cause traps.
146 * When it hits an instruction that requires that the trapped op succeeded
147 * (usually because it reads a reg. that the trapped op wrote) then it
148 * causes this exception. We need to emulate all the insns on the queue
149 * and then allow the op to proceed.
150 * This code should also handle the case where the trap was precise,
151 * in which case the queue length is zero and regs->pc points at the
152 * single FPop to be emulated. (this case is untested, though :->)
153 * You'll need this case if you want to be able to emulate all FPops
154 * because the FPU either doesn't exist or has been software-disabled.
155 * [The UltraSPARC makes FP a precise trap; this isn't as stupid as it
156 * might sound because the Ultra does funky things with a superscalar
157 * architecture.]
158 */
159
160 /* You wouldn't believe how often I typed 'ftp' when I meant 'fpt' :-> */
161
162 int i;
163 int retcode = 0; /* assume all succeed */
164 unsigned long insn;
165
166 #ifdef DEBUG_MATHEMU
167 printk("In do_mathemu()... pc is %08lx\n", regs->pc);
168 printk("fpqdepth is %ld\n", fpt->thread.fpqdepth);
169 for (i = 0; i < fpt->thread.fpqdepth; i++)
170 printk("%d: %08lx at %08lx\n", i, fpt->thread.fpqueue[i].insn,
171 (unsigned long)fpt->thread.fpqueue[i].insn_addr);
172 #endif
173
174 if (fpt->thread.fpqdepth == 0) { /* no queue, guilty insn is at regs->pc */
175 #ifdef DEBUG_MATHEMU
176 printk("precise trap at %08lx\n", regs->pc);
177 #endif
178 if (!get_user(insn, (u32 *)regs->pc)) {
179 retcode = do_one_mathemu(insn, &fpt->thread.fsr, fpt->thread.float_regs);
180 if (retcode) {
181 /* in this case we need to fix up PC & nPC */
182 regs->pc = regs->npc;
183 regs->npc += 4;
184 }
185 }
186 return retcode;
187 }
188
189 /* Normal case: need to empty the queue... */
190 for (i = 0; i < fpt->thread.fpqdepth; i++) {
191 retcode = do_one_mathemu(fpt->thread.fpqueue[i].insn, &(fpt->thread.fsr), fpt->thread.float_regs);
192 if (!retcode) /* insn failed, no point doing any more */
193 break;
194 }
195 /* Now empty the queue and clear the queue_not_empty flag */
196 if(retcode)
197 fpt->thread.fsr &= ~(0x3000 | FSR_CEXC_MASK);
198 else
199 fpt->thread.fsr &= ~0x3000;
200 fpt->thread.fpqdepth = 0;
201
202 return retcode;
203 }
204
205 /* All routines returning an exception to raise should detect
206 * such exceptions _before_ rounding to be consistant with
207 * the behavior of the hardware in the implemented cases
208 * (and thus with the recommendations in the V9 architecture
209 * manual).
210 *
211 * We return 0 if a SIGFPE should be sent, 1 otherwise.
212 */
213 static inline int record_exception(unsigned long *pfsr, int eflag)
214 {
215 unsigned long fsr = *pfsr;
216 int would_trap;
217
218 /* Determine if this exception would have generated a trap. */
219 would_trap = (fsr & ((long)eflag << FSR_TEM_SHIFT)) != 0UL;
220
221 /* If trapping, we only want to signal one bit. */
222 if(would_trap != 0) {
223 eflag &= ((fsr & FSR_TEM_MASK) >> FSR_TEM_SHIFT);
224 if((eflag & (eflag - 1)) != 0) {
225 if(eflag & FP_EX_INVALID)
226 eflag = FP_EX_INVALID;
227 else if(eflag & FP_EX_OVERFLOW)
228 eflag = FP_EX_OVERFLOW;
229 else if(eflag & FP_EX_UNDERFLOW)
230 eflag = FP_EX_UNDERFLOW;
231 else if(eflag & FP_EX_DIVZERO)
232 eflag = FP_EX_DIVZERO;
233 else if(eflag & FP_EX_INEXACT)
234 eflag = FP_EX_INEXACT;
235 }
236 }
237
238 /* Set CEXC, here is the rule:
239 *
240 * In general all FPU ops will set one and only one
241 * bit in the CEXC field, this is always the case
242 * when the IEEE exception trap is enabled in TEM.
243 */
244 fsr &= ~(FSR_CEXC_MASK);
245 fsr |= ((long)eflag << FSR_CEXC_SHIFT);
246
247 /* Set the AEXC field, rule is:
248 *
249 * If a trap would not be generated, the
250 * CEXC just generated is OR'd into the
251 * existing value of AEXC.
252 */
253 if(would_trap == 0)
254 fsr |= ((long)eflag << FSR_AEXC_SHIFT);
255
256 /* If trapping, indicate fault trap type IEEE. */
257 if(would_trap != 0)
258 fsr |= (1UL << 14);
259
260 *pfsr = fsr;
261
262 return (would_trap ? 0 : 1);
263 }
264
265 typedef union {
266 u32 s;
267 u64 d;
268 u64 q[2];
269 } *argp;
270
271 static int do_one_mathemu(u32 insn, unsigned long *pfsr, unsigned long *fregs)
272 {
273 /* Emulate the given insn, updating fsr and fregs appropriately. */
274 int type = 0;
275 /* r is rd, b is rs2 and a is rs1. The *u arg tells
276 whether the argument should be packed/unpacked (0 - do not unpack/pack, 1 - unpack/pack)
277 non-u args tells the size of the argument (0 - no argument, 1 - single, 2 - double, 3 - quad */
278 #define TYPE(dummy, r, ru, b, bu, a, au) type = (au << 2) | (a << 0) | (bu << 5) | (b << 3) | (ru << 8) | (r << 6)
279 int freg;
280 argp rs1 = NULL, rs2 = NULL, rd = NULL;
281 FP_DECL_EX;
282 FP_DECL_S(SA); FP_DECL_S(SB); FP_DECL_S(SR);
283 FP_DECL_D(DA); FP_DECL_D(DB); FP_DECL_D(DR);
284 FP_DECL_Q(QA); FP_DECL_Q(QB); FP_DECL_Q(QR);
285 int IR;
286 long fsr;
287
288 #ifdef DEBUG_MATHEMU
289 printk("In do_mathemu(), emulating %08lx\n", insn);
290 #endif
291
292 if ((insn & 0xc1f80000) == 0x81a00000) /* FPOP1 */ {
293 switch ((insn >> 5) & 0x1ff) {
294 case FSQRTQ: TYPE(3,3,1,3,1,0,0); break;
295 case FADDQ:
296 case FSUBQ:
297 case FMULQ:
298 case FDIVQ: TYPE(3,3,1,3,1,3,1); break;
299 case FDMULQ: TYPE(3,3,1,2,1,2,1); break;
300 case FQTOS: TYPE(3,1,1,3,1,0,0); break;
301 case FQTOD: TYPE(3,2,1,3,1,0,0); break;
302 case FITOQ: TYPE(3,3,1,1,0,0,0); break;
303 case FSTOQ: TYPE(3,3,1,1,1,0,0); break;
304 case FDTOQ: TYPE(3,3,1,2,1,0,0); break;
305 case FQTOI: TYPE(3,1,0,3,1,0,0); break;
306 case FSQRTS: TYPE(2,1,1,1,1,0,0); break;
307 case FSQRTD: TYPE(2,2,1,2,1,0,0); break;
308 case FADDD:
309 case FSUBD:
310 case FMULD:
311 case FDIVD: TYPE(2,2,1,2,1,2,1); break;
312 case FADDS:
313 case FSUBS:
314 case FMULS:
315 case FDIVS: TYPE(2,1,1,1,1,1,1); break;
316 case FSMULD: TYPE(2,2,1,1,1,1,1); break;
317 case FDTOS: TYPE(2,1,1,2,1,0,0); break;
318 case FSTOD: TYPE(2,2,1,1,1,0,0); break;
319 case FSTOI: TYPE(2,1,0,1,1,0,0); break;
320 case FDTOI: TYPE(2,1,0,2,1,0,0); break;
321 case FITOS: TYPE(2,1,1,1,0,0,0); break;
322 case FITOD: TYPE(2,2,1,1,0,0,0); break;
323 case FMOVS:
324 case FABSS:
325 case FNEGS: TYPE(2,1,0,1,0,0,0); break;
326 default:
327 #ifdef DEBUG_MATHEMU
328 printk("unknown FPop1: %03lx\n",(insn>>5)&0x1ff);
329 #endif
330 }
331 } else if ((insn & 0xc1f80000) == 0x81a80000) /* FPOP2 */ {
332 switch ((insn >> 5) & 0x1ff) {
333 case FCMPS: TYPE(3,0,0,1,1,1,1); break;
334 case FCMPES: TYPE(3,0,0,1,1,1,1); break;
335 case FCMPD: TYPE(3,0,0,2,1,2,1); break;
336 case FCMPED: TYPE(3,0,0,2,1,2,1); break;
337 case FCMPQ: TYPE(3,0,0,3,1,3,1); break;
338 case FCMPEQ: TYPE(3,0,0,3,1,3,1); break;
339 default:
340 #ifdef DEBUG_MATHEMU
341 printk("unknown FPop2: %03lx\n",(insn>>5)&0x1ff);
342 #endif
343 }
344 }
345
346 if (!type) { /* oops, didn't recognise that FPop */
347 #ifdef DEBUG_MATHEMU
348 printk("attempt to emulate unrecognised FPop!\n");
349 #endif
350 return 0;
351 }
352
353 /* Decode the registers to be used */
354 freg = (*pfsr >> 14) & 0xf;
355
356 *pfsr &= ~0x1c000; /* clear the traptype bits */
357
358 freg = ((insn >> 14) & 0x1f);
359 switch (type & 0x3) { /* is rs1 single, double or quad? */
360 case 3:
361 if (freg & 3) { /* quadwords must have bits 4&5 of the */
362 /* encoded reg. number set to zero. */
363 *pfsr |= (6 << 14);
364 return 0; /* simulate invalid_fp_register exception */
365 }
366 /* fall through */
367 case 2:
368 if (freg & 1) { /* doublewords must have bit 5 zeroed */
369 *pfsr |= (6 << 14);
370 return 0;
371 }
372 }
373 rs1 = (argp)&fregs[freg];
374 switch (type & 0x7) {
375 case 7: FP_UNPACK_QP (QA, rs1); break;
376 case 6: FP_UNPACK_DP (DA, rs1); break;
377 case 5: FP_UNPACK_SP (SA, rs1); break;
378 }
379 freg = (insn & 0x1f);
380 switch ((type >> 3) & 0x3) { /* same again for rs2 */
381 case 3:
382 if (freg & 3) { /* quadwords must have bits 4&5 of the */
383 /* encoded reg. number set to zero. */
384 *pfsr |= (6 << 14);
385 return 0; /* simulate invalid_fp_register exception */
386 }
387 /* fall through */
388 case 2:
389 if (freg & 1) { /* doublewords must have bit 5 zeroed */
390 *pfsr |= (6 << 14);
391 return 0;
392 }
393 }
394 rs2 = (argp)&fregs[freg];
395 switch ((type >> 3) & 0x7) {
396 case 7: FP_UNPACK_QP (QB, rs2); break;
397 case 6: FP_UNPACK_DP (DB, rs2); break;
398 case 5: FP_UNPACK_SP (SB, rs2); break;
399 }
400 freg = ((insn >> 25) & 0x1f);
401 switch ((type >> 6) & 0x3) { /* and finally rd. This one's a bit different */
402 case 0: /* dest is fcc. (this must be FCMPQ or FCMPEQ) */
403 if (freg) { /* V8 has only one set of condition codes, so */
404 /* anything but 0 in the rd field is an error */
405 *pfsr |= (6 << 14); /* (should probably flag as invalid opcode */
406 return 0; /* but SIGFPE will do :-> ) */
407 }
408 break;
409 case 3:
410 if (freg & 3) { /* quadwords must have bits 4&5 of the */
411 /* encoded reg. number set to zero. */
412 *pfsr |= (6 << 14);
413 return 0; /* simulate invalid_fp_register exception */
414 }
415 /* fall through */
416 case 2:
417 if (freg & 1) { /* doublewords must have bit 5 zeroed */
418 *pfsr |= (6 << 14);
419 return 0;
420 }
421 /* fall through */
422 case 1:
423 rd = (void *)&fregs[freg];
424 break;
425 }
426 #ifdef DEBUG_MATHEMU
427 printk("executing insn...\n");
428 #endif
429 /* do the Right Thing */
430 switch ((insn >> 5) & 0x1ff) {
431 /* + */
432 case FADDS: FP_ADD_S (SR, SA, SB); break;
433 case FADDD: FP_ADD_D (DR, DA, DB); break;
434 case FADDQ: FP_ADD_Q (QR, QA, QB); break;
435 /* - */
436 case FSUBS: FP_SUB_S (SR, SA, SB); break;
437 case FSUBD: FP_SUB_D (DR, DA, DB); break;
438 case FSUBQ: FP_SUB_Q (QR, QA, QB); break;
439 /* * */
440 case FMULS: FP_MUL_S (SR, SA, SB); break;
441 case FSMULD: FP_CONV (D, S, 2, 1, DA, SA);
442 FP_CONV (D, S, 2, 1, DB, SB);
443 case FMULD: FP_MUL_D (DR, DA, DB); break;
444 case FDMULQ: FP_CONV (Q, D, 4, 2, QA, DA);
445 FP_CONV (Q, D, 4, 2, QB, DB);
446 case FMULQ: FP_MUL_Q (QR, QA, QB); break;
447 /* / */
448 case FDIVS: FP_DIV_S (SR, SA, SB); break;
449 case FDIVD: FP_DIV_D (DR, DA, DB); break;
450 case FDIVQ: FP_DIV_Q (QR, QA, QB); break;
451 /* sqrt */
452 case FSQRTS: FP_SQRT_S (SR, SB); break;
453 case FSQRTD: FP_SQRT_D (DR, DB); break;
454 case FSQRTQ: FP_SQRT_Q (QR, QB); break;
455 /* mov */
456 case FMOVS: rd->s = rs2->s; break;
457 case FABSS: rd->s = rs2->s & 0x7fffffff; break;
458 case FNEGS: rd->s = rs2->s ^ 0x80000000; break;
459 /* float to int */
460 case FSTOI: FP_TO_INT_S (IR, SB, 32, 1); break;
461 case FDTOI: FP_TO_INT_D (IR, DB, 32, 1); break;
462 case FQTOI: FP_TO_INT_Q (IR, QB, 32, 1); break;
463 /* int to float */
464 case FITOS: IR = rs2->s; FP_FROM_INT_S (SR, IR, 32, int); break;
465 case FITOD: IR = rs2->s; FP_FROM_INT_D (DR, IR, 32, int); break;
466 case FITOQ: IR = rs2->s; FP_FROM_INT_Q (QR, IR, 32, int); break;
467 /* float to float */
468 case FSTOD: FP_CONV (D, S, 2, 1, DR, SB); break;
469 case FSTOQ: FP_CONV (Q, S, 4, 1, QR, SB); break;
470 case FDTOQ: FP_CONV (Q, D, 4, 2, QR, DB); break;
471 case FDTOS: FP_CONV (S, D, 1, 2, SR, DB); break;
472 case FQTOS: FP_CONV (S, Q, 1, 4, SR, QB); break;
473 case FQTOD: FP_CONV (D, Q, 2, 4, DR, QB); break;
474 /* comparison */
475 case FCMPS:
476 case FCMPES:
477 FP_CMP_S(IR, SB, SA, 3);
478 if (IR == 3 &&
479 (((insn >> 5) & 0x1ff) == FCMPES ||
480 FP_ISSIGNAN_S(SA) ||
481 FP_ISSIGNAN_S(SB)))
482 FP_SET_EXCEPTION (FP_EX_INVALID);
483 break;
484 case FCMPD:
485 case FCMPED:
486 FP_CMP_D(IR, DB, DA, 3);
487 if (IR == 3 &&
488 (((insn >> 5) & 0x1ff) == FCMPED ||
489 FP_ISSIGNAN_D(DA) ||
490 FP_ISSIGNAN_D(DB)))
491 FP_SET_EXCEPTION (FP_EX_INVALID);
492 break;
493 case FCMPQ:
494 case FCMPEQ:
495 FP_CMP_Q(IR, QB, QA, 3);
496 if (IR == 3 &&
497 (((insn >> 5) & 0x1ff) == FCMPEQ ||
498 FP_ISSIGNAN_Q(QA) ||
499 FP_ISSIGNAN_Q(QB)))
500 FP_SET_EXCEPTION (FP_EX_INVALID);
501 }
502 if (!FP_INHIBIT_RESULTS) {
503 switch ((type >> 6) & 0x7) {
504 case 0: fsr = *pfsr;
505 if (IR == -1) IR = 2;
506 /* fcc is always fcc0 */
507 fsr &= ~0xc00; fsr |= (IR << 10); break;
508 *pfsr = fsr;
509 break;
510 case 1: rd->s = IR; break;
511 case 5: FP_PACK_SP (rd, SR); break;
512 case 6: FP_PACK_DP (rd, DR); break;
513 case 7: FP_PACK_QP (rd, QR); break;
514 }
515 }
516 if(_fex == 0)
517 return 1; /* success! */
518 return record_exception(pfsr, _fex);
519 }
520